Reinforcement Learning Configuration Interaction.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Electronic Structure Theory

Background:

• Selected configuration interaction (sCI) methods offer computational savings and wave function compression by exploiting the sparsity of full configuration interaction (FCI) wave functions.
• A key challenge in sCI methods is the effective selection of important determinants to include in the calculation.
• Existing sCI techniques face limitations in accurately and efficiently identifying crucial determinants.

Purpose of the Study:

• To explore the application of reinforcement learning (RL) for solving the determinant selection problem in sCI.
• To develop a novel method, termed reinforcement-learned configuration interaction (RLCI), for accurate and efficient wave function compression.
• To demonstrate the potential of RL in addressing complex challenges within electronic structure theory.

Main Methods:

• The configuration interaction problem is reformulated as a sequential decision-making process.
• A reinforcement learning agent is trained to learn which determinants to include or exclude dynamically.
• The RL agent learns on-the-fly, optimizing determinant selection during the computation.

Main Results:

• The developed reinforcement-learned configuration interaction (RLCI) method achieves near-FCI accuracy.
• RLCI yields a compressed wave function with significant computational savings compared to traditional methods.
• The approach effectively addresses the determinant selection challenge in sCI.

Conclusions:

• Reinforcement learning provides a powerful new approach to tackle the determinant selection problem in selected configuration interaction.
• The RLCI method offers a promising tool for accurate and computationally efficient electronic structure calculations.
• This work highlights the broader applicability of reinforcement learning in advancing computational quantum chemistry.